A course where USAAIO Round 1 finishers Round 2 (MIT campus · GPU L4 · "Everything" scope)prepare for. On top of the full syllabus, students put GPU-based model training and advanced topics (Transformers·NLP·CV·Generative AI) into real practice. Advanced A1 and A2 work is required alongside, and we even simulate adapting to an unfamiliar campus environment.
Published: May 16, 2026 | Last updated: May 16, 2026 · Based on official USAAIO information
Track
USAAIO-only
Round 2 (MIT) prep
Target
Round 1 finishers
A1·A2 completed/concurrent
Recommended Hours
About 13 hours
1:1 · varies 8–18 hours
USAAIO mapping
Full Round 2 scope
Transformers~Generative AI
After finishing U3, students can, in a GPU (L4) environment, fine-tune pretrained models and run advanced architectures in practice, and deliver the same performance even in the unfamiliar setting of the MIT campus.
On the foundation built through Round 1, U3 adds Round 2's "Everything" scope(Transformers·NLP·CV·Generative AI), plus knowing when to use the GPU, memory management, and the Authorized Documentation/Allowed Models.
USAAIO Passing Round 1is the prerequisite, and A1·A2 Completion (or concurrent enrollment) is Required. For Round 1 prep, see U2; for the full track, see USAAIO Prep Curriculum Hub reference.
Below is the standard 6-week plan for 1:1. We run intensively from the Round 1 results announcement until right before Round 2, and all hands-on work is standardized on a GPU (L4) environment.
| Week | Topic | Deliverable |
|---|---|---|
| 1 | Transformers deep dive: attention by hand, architecture variants | from-scratch transformer mini |
| 2 | NLP: tokenization, embeddings, pretraining/fine-tuning | BERT fine-tuning Colab notebook |
| 3 | CV deep dive: object detection, UNet | YOLO/UNet training notebook |
| 4 | Generative AI: Autoencoder, VAE, GAN | VAE + GAN training notebook |
| 5 | Stable Diffusion + Denoising Diffusion theory and practice | Diffusion model inference notebook |
| 6 | Round 2 integrated simulation (MIT environment recreated, GPU used) | full mock exam |
※ Weeks are content units, and actual time required varies by student. Recommended about 13 hours, ranging 8–18 hours.
Round 2 allows the Colab L4 GPU. We train students to make using the GPU at the right time, managing memory, and allocating training time second nature.
We offer an option to hold mock exams at an outside location so students adapt to an unfamiliar campus, English instructions, and other test-takers.
Students study the Authorized Documentation and Allowed Models list that Round 2 publishes in advance, and internalize that using anything off the list means disqualification.
We cover the entire Round 2 scope across 6 GPU mock exams, accumulating logs of GPU usage timing, memory, and training time to quantify operating efficiency.
Students build up notebooks for fine-tuning pretrained models and practicing advanced architectures, based on HuggingFace.
Round 2 takes place on the MIT campus. The experience of taking on Round 2 and your HuggingFace notebooks become a very strong signal in U.S. college admissions.
GitHub
an organized repo usaaio-round2-prep
HuggingFace
Collection of fine-tuning and advanced-architecture notebooks
Track record
Experience taking on Round 2 @ MIT (admissions signal)
These deliverables are KOAI C1 Portfolio Studio and reused in U.S. college admissions consulting.
U3 is the intensive Round 2 course for students who passed Round 1. For the full track structure, see USAAIO Prep Curriculum Hub.
Current Course
U3. Round 2 Bootcamp
MIT · GPU L4 · Everything
It's exclusively for students who passed USAAIO Round 1. Completing A1 and A2 (or taking them concurrently) is required, and it covers Round 2's "Everything" scope and GPU-based model training.
Round 2 uses GPU L4 on the MIT campus, and the scope expands to everything (including Transformers·NLP·CV·Generative AI). Using pretrained models and knowing the Allowed Models list are key.
We offer an option to hold mock exams at an outside location so students adapt to an unfamiliar campus, English instructions, and other test-takers. It removes the shock of the exam-day environment in advance.
No. It's a recommended figure for 1:1, varying by ±30–50% (about 8–18 hours). We run intensively from the Round 1 results announcement until right before Round 2, and we work out an individual plan at the first diagnostic.
If you pass Round 2, it leads into U4 (Camp prep). The Round 2 @ MIT experience and your HuggingFace notebooks can also be repurposed as assets for U.S. college admissions. Check the USAAIO competition guide for the schedule.
Based on whether you passed Round 1 and your A1·A2 completion status, we'll individually design your intensive Round 2 schedule in a diagnostic session.